CalmSense

Wearable stress benchmark.

• 15 subjects · 58 saved features · LOSO.
• RF binary accuracy 0.910, averaged across subjects.
• Static results; no significant feature-model differences.

WESAD · LOSO · SHAP · Calibration

Code Demo

PulseShift

Bike-share demand suppression forecast.

• 26,288 panel hours; 2024 unweighted holdout AUROC 0.936, ECE 0.044.
• Open-Meteo/CAMS; exclude heat index ≥103°F or AQI ≥150.
• Individual activity and health benefits unmeasured.

Demand · Weather · AQI · Logistic

Code Demo

NovaVision

Text-to-image emotion pilot.

• 7 classes · 256 px · CPU.
• No conditioning tier passes Holm correction.
• Probe collapse; correction limited to 5/7 classes.

Affect · SD-Turbo · CLIP · Evaluation

Code Video

NexusRAG

Local search and cited answers.

• PDF/DOCX/MD/TXT → LanceDB + BM25 → Ollama.
• PRF once below 0.55 similarity; NLI optional.
• Citation indices checked; claims need review.

RAG · LanceDB · BM25 · Ollama

Code Video

Multimodal-Multisensor

Longitudinal eye-tracking and HRV study.

• 10 adults · 3 weekly sessions.
• ICC(1) 0.22–0.61 across reported measures.
• Reported HRV interval crosses zero; underpowered.

Longitudinal · Eye-tracking · HRV · ICC

Code Demo Paper